User modelling for an adaptive system: An intelligent email client

This research explores problems of email filtering as an application of Machine Learning and User Modelling. The main objective of this thesis is to create a prioritization model that identifies high priority emails. The hypothesis is that this produces a user model that can filter out spam email through the assignment of low priority values to offending messages and rank legitimate emails in terms of their importance for higher user productivity. The secondary objective of this thesis is to create a categorization model that classifies incoming messages to existing email folders. The success of the prioritization model, developed by a Back Propagation neural network, depends on a user's dictionary that provides a unique input feature representing the user's interests. The categorization model is implemented as a Multiple Task Learning neural network and the most probable three folders are recommended to the user. A new method, called k-fold Chronological Cross-Validation (CCV), is proposed that realistically evaluates the effectiveness of email filtering by taking into consideration the temporal nature of email messages. An intelligent email client prototype is developed based on the proposed method. Empirical studies on two large corpora of email messages demonstrated the success of the methods.